Shape inference: GatherBlockQuantized dispatcher (#23748)

### Description
Add shape infer dispatcher for `GatherBlockQuantized` contrib op. It
reuses the dispatcher for `Gather` op since the first two inputs have
the same specs. The output elem type comes from input 2 (scales) for
`GatherBlockQuantized`.

### Motivation and Context
Support shape inference for models with `GatherBlockQuantized` op.

Author: jambayk

onnxruntime/python/tools/symbolic_shape_infer.py

@@ -202,6 +202,7 @@ def __init__(self, int_max, auto_merge, guess_output_rank, verbose, prefix=""):
             "EmbedLayerNormalization": self._infer_EmbedLayerNormalization,
             "FastGelu": self._infer_FastGelu,
             "GatedRelativePositionBias": self._infer_GatedRelativePositionBias,
+            "GatherBlockQuantized": self._infer_Gather,
             "Gelu": self._infer_Gelu,
             "GemmFastGelu": self._infer_GemmFastGelu,
             "GemmFloat8": self._infer_GemmFloat8,
@@ -459,6 +460,7 @@ def _onnx_infer_single_node(self, node):
             "BiasGelu",
             "EmbedLayerNormalization",
             "FastGelu",
+            "GatherBlockQuantized",
             "Gelu",
             "GemmFastGelu",
             "LayerNormalization",
@@ -1118,10 +1120,17 @@ def _infer_Gather(self, node):  # noqa: N802
         axis = handle_negative_axis(get_attribute(node, "axis", 0), len(data_shape))
         indices_shape = self._get_shape(node, 1)
         vi = self.known_vi_[node.output[0]]
+        if node.op_type == "Gather":
+            elem_type = self.known_vi_[node.input[0]].type.tensor_type.elem_type
+        elif node.op_type == "GatherBlockQuantized":
+            # scales
+            elem_type = self.known_vi_[node.input[2]].type.tensor_type.elem_type
+        else:
+            raise ValueError(f"Unsupported Gather op_type: {node.op_type}")
         vi.CopyFrom(
             helper.make_tensor_value_info(
                 node.output[0],
-                self.known_vi_[node.input[0]].type.tensor_type.elem_type,
+                elem_type,
                 data_shape[:axis] + indices_shape + data_shape[axis + 1 :],
             )
         )